Leonardo.Aiさんのイラストまとめ

Tanky crab man is an interesting one - our initial prompting gave some hilarious results 😂

Paring the prompt back and upping the init strength got us much closer to the source material, and finishing can be done over in our canvas tooling (coming soon!)

https://t.co/OOHsdaQSC2

We're surprised by how well it performs, and interestingly - the order in which the training data was organised had a significant impact on the final model.

Another training run on the same dataset in a different order didn't garner anywhere near the same consistency of results.

We've also built in some really fun tools like full #gpt3 prompt ideation and auto-batching against prompts, so you can leave a job running and come back a few minutes later to hundreds of generations.

While our platform has a core focus on game asset production, it offers so much more.

We trained a model on #MJ4 outputs to see how well we could improve the baseline #StableDiffusion model. Quite a lot it turns out! 🔥

Signup for Early Access - https://t.co/OOHsdb91Qa

🧵👇

We then tried "a scientist", "a president", "a soldier", "an astronaut".

Boom. Boom. Boom. Boom. 🤯

All perfectly usable. By the way, background removal is automated on Leonardo, we left them in here as they look nicer for the Tweet 🙂

Let's try a prompt that's nowhere close to the dataset, how about "a zombie". 🧟‍♂️

This is interesting - the lower strength training (left) gives you a different feel with the pixel art and can deviate further stylistically, while the stronger runs (right) look more polished.

Okay all the hard work is out of the way. Let's get to generating.

"a humble farmer" - the images in this tweet are from one of the stronger training runs.

🤔 What about something a little weirder. 👇

Or are you an artist looking to supercharge your workflow? We can do that too.

The possibilities are limitless.

Want a grain silo? ☑️
Particle accelerator? ☑️
Gaming icons? ☑️
UI elements? ☑️
Pretty much anything at all? ☑️

All of this high quality, quickly, and at scale.

Generate production-ready assets like these at scale, using pre-trained models, or train your own.

With clever tooling and workflows to supercharge productivity and creativity, we can't wait to ship what we've been working on. 😎

This thread is To Be Continued...

However, we promised dragons, so here's a teaser 🐉

Remember to hit us up in DMs if you're keen to join the alpha. Until next time!